📚 English Word Annotator

The name sounds fancier than the job, and that's the point. 📚 English Word Annotator doesn't detect anything, call any API, or need a key. It's a pure renderer: you hand it an image, a list of bounding boxes, and a list of English labels, and it draws a colored rectangle around each object with the word printed in a little colored tag. It was built for one specific thing - vocabulary learning cards - and it's honestly the cleanest way to get there in ComfyUI.

Here's the workflow it's meant to sit in. Generate a scene (SDXL, Flux, whatever - the README demos Z-Image-Turbo), then run Florence-2 in detection mode to get per-object boxes and labels, then feed both into this node and save the result. The rendering step is the part nobody wants to hand-roll in PIL, and that's the whole package: one node, one file, about 250 lines.

How it works

The mechanism is boring in the best way. The image input (an IMAGE tensor) gets converted to a PIL image, boxes are drawn with ImageDraw, labels are stamped in a filled pill above or inside each box, and the result goes back to an IMAGE tensor. Two details save you from real pain:

• An AnyType wildcard on bboxes and labels. Florence-2 wrappers have changed their output shape across releases - flat lists, batch-wrapped lists, even {"bbox":..., "label":...} dicts from older versions. This node normalizes all of them. If you're fighting a type mismatch between Florence-2 versions, this is the node that shrugs and works.
• Sensible fallbacks. Boxes are clipped to image bounds, degenerate boxes are skipped, and if a label tag doesn't fit above the box it falls back inside. If you're missing a label for a box, you get #1, #2… instead of a crash.

An 8-color palette cycles so adjacent boxes stay distinguishable - no two neighbors get the same tint.

The inputs that matter

Required: image, bboxes, labels. That's it - everything else has a sensible default.

• label_position (above / inside_top / below) - where the word tag sits. above is the default and usually the right call for study cards; inside_top if objects run to the top edge.
• font_size (default 36) and box_width (default 4) - text size and outline thickness. Bump font_size for big classroom cards.
• max_labels (default 0, and 0 = no limit) - cap how many objects get annotated. Handy when Florence-2 finds 20 regions and you only want the first five.
• font_path and text_color (both optional) - an explicit TTF path if you hate the defaults, and a hex text color.

One output: image. Wire it straight into SaveImage or PreviewImage. It only draws on the first frame of a batch, so don't feed it a video.

Install

It's two minutes. Either ComfyUI Manager (search "English Annotator") or:

cd ComfyUI/custom_nodes
git clone https://github.com/aadebuger/ComfyUI-EnglishAnnotator.git
cd ComfyUI-EnglishAnnotator
uv pip install -r requirements.txt   # just Pillow + numpy

Restart ComfyUI and it appears under 📚 English Annotator. The node itself is featherweight - the real dependency is upstream: you'll also want kijai/ComfyUI-Florence2 (and its Florence-2-base/large weights, which download on first run and can be several GB). The README's example flow is CheckpointLoader → KSampler → VAEDecode feeding the image, with Florence2ModelLoader → Florence2Run feeding boxes and labels.

Gotchas

• Labels render as empty boxes or squares on a fresh headless Linux box - no fonts installed. sudo apt install fonts-dejavu-core fixes it.
• Florence-2 is the fragile half. Its transformers dependency needs >=4.41 but can clash with other packs locked to 4.x; if Florence2ModelLoader complains, upgrade it but stay below 5.0. And flash_attn import errors - set attention="sdpa" on Florence2Run.
• Long labels overflow their colored pill - the box is clamped to the image edge but the text isn't wrapped, so a word like "refrigerator" on a small box will spill right. Keep words short or bump font_size down.
• Uses uv pip install, not pip if your venv is uv-managed - uv venvs ship no pip, and python -m pip will just fail.

For task mode, od gives clean COCO category names; caption_to_phrase_grounding is the one to reach for when you have a fixed vocabulary list and want Florence-2 to lock onto exactly those words - which is precisely the daily-vocab pipeline this node was built for.